em driver sending bad packet lengths

[Kris Kennaway]

I am seeing the em driver on 7.0 sending packets with bad UDP (and 
apparently sometimes IP) packet length fields.  This is from UDP NFS 
traffic (90MB/sec over gige).  These packets are dropped on reception by 
the kernel and counted.

Here is a tcpdump trace showing some bad packets:

17:50:27.912629 IP bad-len 0
17:50:27.912748 IP bad-len 0
17:50:27.912764 IP bad-len 0

I havent yet caught a better tcpdump of the UDP bad packets but they are 
logged by netstat -s

hydra1# netstat -s -p udp
udp:
         120091992 datagrams received
         0 with incomplete header
         1775 with bad data length field
         ^^^^
         14840 with bad checksum
         0 with no checksum
         16 dropped due to no socket
         258 broadcast/multicast datagrams undelivered
         0 dropped due to full socket buffers
         0 not for hashed pcb
         120075103 delivered
         120190737 datagrams output
         0 times multicast source filter matched

Disabling txcsum and rxcsum didnt help.

em0: <Intel(R) PRO/1000 Network Connection Version - 6.5.3> port 
0x2000-0x201f mem 0xd8a00000-0xd8a1ffff irq 18 at device 0.0 on pci4
em0: Ethernet address: 00:30:48:33:54:ca
em0: [FILTER]

em0 at pci4:0:0:   class=0x020000 card=0x000015d9 chip=0x10968086 rev=0x01 
hdr=0x00
     vendor     = 'Intel Corporation'
     device     = 'PRO/1000 EB Network Connection'
     class      = network
     subclass   = ethernet

em0: <Intel(R) PRO/1000 Network Connection Version - 6.5.3> port 
0x4000-0x401f mem 0xe1000000-0xe101ffff irq 16 at device 0.0 on pci18
em0: Ethernet address: 00:30:48:8f:55:48
em0: [FILTER]

em0 at pci18:0:0:  class=0x020000 card=0x108c15d9 chip=0x108c8086 rev=0x03 
hdr=0x00
     vendor     = 'Intel Corporation'
     device     = 'PRO/1000 PM'
     class      = network
     subclass   = ethernet

The second-order effect from this is that our NFS client currently 
misbehaves when UDP packets are dropped, and this leads to data corruption.

Kris